INPUT TEXT

(English Caption from OCR)

TOKENIZATION & LANGUAGE TAGGING

(en_XX source language token inserted)

ENCODER (12 Layers)

Multi-Head Attention Feed-Forward Network

1 Residual + LayerNorm applied at every block

Encoded Multilingual Representation

TARGET LANGUAGE FORCED BOS TOKEN

(ta_IN / te_IN / ml_IN during decoding)

DECODER (12 Layers)

Self-Attention (Masked) Cross-Attention (Encoder)

+ Feed-Forward Network

Residual Connections + Layer Normalization

LINEAR + SOFTMAX LAYERS

(Generate Next Token Probability)

TRANSLATED OUTPUT SENTENCE

(Tamil / Telugu / Malayalam Caption)